CN103310466B - A kind of monotrack method and implement device thereof - Google Patents

Abstract

A kind of monotrack method and implement device thereof, comprising: at the video V={F of N frame gray level image composition ₀, F ₁, F _nin, at F ₀selected target O in frame ₀, to image O ₀carry out gray processing process and wide high normalized, obtain the initiation parameter of this image; Sorter initialization and renewal thereof, comprise the structure of training set, feature extraction and model modification three part; Tracking target, utilizes model f _t+1at F _t+1two field picture carries out target following.The present invention is based on the compressed sensing dimension reduction method of binary feature to express the apparent of target, can the deformation of effective expression target, improve anti-blocking and the ability of illumination, thus can robust tracking target, there is the advantage that memory consumption is low and calculated amount is little simultaneously, reach real-time follow-up speed.

Description

A kind of monotrack method and implement device thereof

Technical field

The present invention relates to the tracking of the reference position that sets the goal and implement device thereof, especially a kind of monotrack method and implement device thereof.

Background technology

Carrying out the research of tracking technique based on moving target feature, is the focus of computer vision field research in recent years.Study widely although the biological characteristics such as fingerprint, palmmprint and vein have been carried out in security fields and had Preliminary Applications, these biological characteristics belong to contact identification, greatly limit its range of application.Comparatively speaking, gait and recognition of face this " contactless " recognition technology, carry out ingenious combination by human motion and biological characteristic and carry out Study of recognition, become a key areas in intelligent scene video monitoring at present.Especially Gait Recognition, needs the feature collecting movement human when being expert at people's walking movement to carry out identification, and the accuracy of the movement human detection and tracking work in early stage, real-time are the prerequisites of overall recognition performance.This proposes very large challenge to video monitoring, and based on the requirement of system scenarios security performance, the manually-operated video monitoring of traditional dependence, owing to having following shortcoming, can not adapt to the needs of actual scene security monitoring application.Under the complex background of reality, how the key of intelligent video monitoring system is not only in the tracking of realize target, and in intelligent transportation or man-machine interaction application, also have very important effect, therefore target tracking algorism obtains and develops very widely.But the success or not of most of pedestrian tracking algorithm, all will depend on the complexity of background and the similarity of pedestrian target and background, just can obtain good result when only having object and background to distinguish larger in color.In order to solve the problem of pedestrian tracking in complex scene, needing us to design increasing robust algorithm, to make in all practical applications such as it is enough to solve illumination variation, noise effect, barrier block inevitably problem.How accurately and rapidly from video sequence detection and tracking to go out moving target extremely important, be one of the most key technology of identification and abnormal behaviour identification, the method for current motion target tracking mainly contains two kinds: 1, statistical learning method; 2, based on the algorithm of color characteristic.First method becomes one of mainstream technology in area of pattern recognition gradually, and it has successful application on many classical problems, and motion target tracking is exactly an example.Adaboost algorithm is a kind of cascade track algorithm that the people such as Freund propose, and its target from Weak Classifier space, automatically picks out several Weak Classifiers be integrated into a strong classifier.The Adaboost algorithm based on Haar type feature that the people such as Vila propose is the successful Application of Adaboost algorithm on Face datection.The people such as Grabner propose online Adaboost algorithm, and Adaboost algorithm is applied to target tracking domain, achieve good tracking effect.Be different from off-line Adaboost algorithm, the training sample of online Adaboost algorithm is one or several data obtained in real time.Use this algorithm can adapt to the problems such as moving target changing features better, but online Adaboost algorithm relies on merely sorter to follow the tracks of, easy classification error river occurrence of large-area is blocked in the background of complexity, cause to follow the tracks of and lose.

The Camshift track algorithm that Bradski proposes relies on its performance good in real-time and robustness to be also a kind of by the algorithm of extensive concern.Camshift algorithm with Meanshift algorithm for core, solve the shortcoming that Meanshift can not change tracking window size, reduce target search scope, improve accuracy and operation efficiency, good tracking effect can be obtained in the simple situation of background.But Camshift track algorithm affects by other moving targets of surrounding comparatively large, easily thinks non-targeted point by mistake impact point, target size is changed and causes to follow the tracks of to lose efficacy, and then occur following the tracks of Loss.Traditional Camshift target tracking algorism colouring information is followed the tracks of as feature, when color of object and background or non-targeted close, also there will be to follow the tracks of and lose.Further, traditional Camshift target tracking algorism easily follows the tracks of failure to fast-moving target, and cannot restore from failure.In view of single online Adaboost algorithm and Camshift algorithm all can not obtain good tracking effect, patent of invention CN201210487250.3, patent name is a kind of motion target tracking method combined based on online Adaboost algorithm and Camshift algorithm for " a kind of motion target tracking method " discloses, first the eigenmatrix of online Adaboost track algorithm and sorter computing are obtained confidence map, local direction histogram feature and color characteristic have been merged in choosing of feature, then on confidence map, Camshift algorithm is applied, make the Fusion Features of Camshift algorithm application texture and colouring information.The method comprises the following steps: the first step, and the fast-moving target detection method based on code book model accurately detects moving target; Second step, to the initialization of online Adaboost algorithm Weak Classifier group, obtain strong classifier, local direction histogram feature and color characteristic have been merged in choosing of moving target feature; 3rd step, the eigenmatrix of online Adaboost track algorithm and Weak Classifier computing are obtained confidence map, confidence map is applied Camshift track algorithm, upgrades Weak Classifier according to the moving target position obtained, finally obtain the tracking results of whole section of video sequence.The method utilizes conventional approach to solve tracking problem, and its existing problems have two aspects, and first, the feature robustness of extraction is not enough, for direction histogram locally and color characteristic often comparatively responsive to noise, thus the robustness causing target apparent is not enough; Secondly, the Camshift method that the method adopts is to illumination, and the target of color change easily produces drift phenomenon, thus result in the problem of tracking accuracy decline, cannot meet the severe rugged environment in actual monitored video.Therefore the two large key issues that the strong sorter of the characteristic sum generalization ability of robust is target following are designed.

Patent name is " the monotrack method based on composing power least square " (publication number: 103093482A, publication date: 2013-05-08) disclose a kind of monotrack method based on composing power least square, in the mode of reconstructed error, target is followed the tracks of.The shortcoming of the method consumes the regular hour at sparse needs thing solve reconstruct, therefore follows the tracks of the tracking that more difficult realization is real-time.Suppose the known video V={F be made up of N frame gray level image ₀, F ₁... F _n, the wide height of two field picture is respectively w, h.The present invention want solve problem be: at F ₀selected target O in frame ₀, then a kind of tracking is proposed to O ₀carry out the tracking of N continuous frame.

This to the tracking of the reference position that sets the goal, the major technique adopted at present is the color by extracting target, and profile and texture feature information express the apparent of target, then utilize the apparent model of classification learning method learning objective; Then in next frame picture, the position of target is detected by apparent model, or by simple track algorithm, as average drifting or light stream etc., the position that tracking target occurs in next frame picture, the result after this integrating tracking and target detection two kinds of methods obtains a tracing positional the most believable; Finally by certain update strategy, adaptive renewal is carried out to apparent model.

Summary of the invention

For solving the problem, the object of the present invention is to provide a kind of by compressed sensing feature extracting method with at line core study update method and device, overcome the deficiency of above-mentioned tracking technique, the compressed sensing feature extremely strong to Objective extraction robustness, improve the ability to express that target is apparent, then just go to upgrade apparent model in line core study way.

For achieving the above object, technical scheme of the present invention is:

A kind of monotrack method, comprising:

The first step, initiation parameter, that is, at F ₀target O is obtained in frame ₀the rectangle frame B of initial position ₀=[x ₀, y ₀, w ₀, h ₀], represent the upper left corner horizontal ordinate of frame respectively, upper left corner ordinate, frame is wide, and frame is high; On the image block I that wide height is identical, generate L random point to collection wherein represent respectively l point to the horizontal ordinate of first point, the ordinate of first point, the horizontal ordinate of second point, the ordinate of second point, the right generating mode of random point is limited in level or vertical two kinds; Generate sparse stochastic matrix A, for Feature Dimension Reduction;

Second step, sorter initialization and renewal thereof, carry out t (t=0,1 ..., N-1) and secondary iteration renewal, by process t two field picture F _t, comprise training sample set D _tstructure, feature extraction and model modification three processing procedures;

3rd step, target following, utilizes model f _t+1at F _t+1two field picture carries out target following, tracking step comprises: forecast sample constructs, feature extraction, sample classification, selects degree of confidence the most much higher sample, generates final (while be also a best) object boundary frame, output tracking frame, t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return second step.

In the first step, A=[α _ij] _{h × L}, line number is H, and value is 50-300, (being preferably 100), and columns is L, there will be a known an equiprobability function rand, it generates equiprobably 1,2,3 ..., an element in 2024}, if rand ∈ 1,2,3 ..., 16}, then if rand ∈ 17,18,19 ..., 32}, then otherwise a _ij=0.

In second step,

Described training sample set D _tstructure comprises the steps:

A) positive sample collection from object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_t^{pos}=\left\{(x,y)\right|x_t-10<x<x_t+10,y_t-10<y<y_t+10\}$ In random extract 50-500 (preferably 100) positive samples pictures collection acquisition methods is Pan and Zoom, and step is as follows:

I. the generation formula of positive sample boundary frame:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(1)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii. carry out 50-500 time (preferably 100 times) to operate as follows: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (1) calculates sample is substituted into; According to bounding box [x', y', w', h'] cut-away view as F _tsubimage I _i; By I _ibe normalized to the image I of wide height identical (such as 32x32) _i; After so carrying out 50-500 time, generate 50-500 and open positive samples pictures set, be designated as

B) negative sample collection ? outer peripheral areas, definitely $Q_t^{neg}=\left\{(x,y)\right|0\&le;x,x\&GreaterEqual;x_t-10,0\&le;y,y\&GreaterEqual;y_t+10\},$ Random acquisition 50-1000 (preferably 200) negative sample collection acquisition methods is translation or convergent-divergent, and step is as follows:

I. the formula of negative sample bounding box is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(2)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, span [0,20]; X, y span is

Ii. carry out 50-1000 time (preferably 200 times) to operate as follows: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (2) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I; I is normalized to the image I of wide height identical (such as 32x32); After so carrying out 50-1000 time, generate 50-1000 and open negative sample picture, be designated as

C) positive negative sample is merged, composing training sample D _t; Definitely, wherein y _i{-1, the 1} class label representing sample ,-1 represents negative sample to ∈, and 1 represents positive sample.

In second step, described feature extraction is for extracting training sample set D _tin the feature of all sample images, extract sample { I _i, y _ithe step of feature is as follows:

A) initialization sample { I _i, y _ifeature be characteristic length is the element number of CP, i.e. L;

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q);

C) sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is 50-300 (being preferably 100), thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

Structure obtains training sample set D thus _tfeature set

In second step, described model modification is for utilizing training sample set D _tfeature set upgrade sorter model wherein i.e. Renewal model parameter w _t∈ R ^{1 × 101}, R represents real number, and step is as follows:

If a) t=0, initialization w _t∈ R ^{1 × 101}, λ value is 0.0001; Otherwise, perform step b);

B) following iterative step is carried out:

I. from Z _ta middle Stochastic choice k sample, forms subset

Ii. from A _tmiddle searching meets the sample set of certain condition

Iii. calculating parameter value η _t=1/ (λ t).

Iv. first time undated parameter:

$w_{t+\frac{1}{2}}=(1-{\mathrm{\&eta;}}_t\mathrm{\&lambda;})w_t+\frac{\mathrm{\&eta;}t}{k}\underset{\{z,y\}\&Element;A_t^{+}}{\&Sigma;}yz$

V. second time undated parameter:

$w_{t+1}=min\left\{1,\frac{1/\sqrt{\mathrm{\&lambda;}}}{\left|\right|w_{t+\frac{1}{2}}\left|\right|}\right\}{w}_{t+\frac{1}{2}}$

Wherein min represents the minimum value in element, || || represent 2 norms.

C) w is exported _t+1.

In 3rd step, describedly utilize model f _t+1at F _t+1the step that two field picture carries out target following is as follows:

1) sample set extract.From object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_{t+1}^u=\left\{(x,y)\right|x_t-30<x<x_t+30,y_t-30<y<y_t+30\}$ In random extract 150-300 (preferably 200) positive samples pictures collection acquisition methods is translation or convergent-divergent, and step is as follows:

I. the formula of sample boundary frame is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(3)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, span [0,20]; X, y span is

Ii. carry out 150-300 time (preferably 200 times) to operate as follows: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (3) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _t+1subimage I; I is normalized to the image I of wide height identical (such as 32x32);

Iii. according to step I i, generate 150-300 and open (preferably 200) samples pictures set, be designated as $D_{t+1}^U={\left\{{\{I_i,1\}}_i\right\}}_{i=1}^{200};$ Sample set sample frame set is in the picture designated as $B_{t+1}^U={\left\{{\{x_i,y_i,w_i,h_i\}}_i\right\}}_{i=1}^{200};$

2) calculate in the feature of every pictures, calculate sample { I _i, y _ithe method of feature is as follows:

A) initialization sample { I _i, y _ifeature be characteristic length is the element number of CP, i.e. L;

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q);

C) sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is 50-300 (being preferably 100), thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

The training sample constructed thus is designated as

As above, sample set to be sorted is formed

3) model f is utilized _t+1to U _t+1in all samples classify, each sample z _i∈ U _t+1produce corresponding degree of confidence:

${\mathrm{Conf}}_i=f_{t+1}\left({\hat{z}}_i\right)=w_{t+1}{\hat{z}}_i$

Wherein degree of confidence is designated as Conf _t+1={ conf ₁, conf ₂..., conf ₂₀₀;

4) according to Conf _t+1, from the most much higher several bounding box of middle selection degree of confidence this number is preferably 1/20 of sample number, generates a final object boundary frame B _t+1=[x _t+1, y _t+1, w _t+1, h _t+1];

5) t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return second step above.

In order to realize said method, the present invention also provides a kind of implement device of monotrack method, comprising:

Image acquiring device, for obtaining a two field picture and carrying out gray processing process and wide high normalized to image from video;

Sorter initialization and updating device thereof, for initialization model and online updating model;

Target tracker, for doing target homing on a new images, makes search result as far as possible consistent with target.

Described image acquiring device comprises random point to generation unit and sparse stochastic matrix generation unit.

Described sorter initialization and updating device thereof comprise samples pictures collection tectonic element, feature extraction unit and model modification unit, and wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing for aligning negative sample image; Model modification unit, the feature samples collection utilizing feature extraction unit to obtain is to upgrade sorter model.

Described target tracker comprises samples pictures collection tectonic element, feature extraction unit and target tracking unit, and wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing, target tracking unit for aligning negative sample image, for classifying to all samples, and therefrom select degree of confidence the most much higher bounding box, and then cluster generates a best object boundary frame.

In sum, a kind of monotrack method provided by the invention is creationary proposes feature extracting method and feature dimension reduction method, the compressed sensing feature extremely strong to Objective extraction robustness, improve the ability to express that target is apparent, then the step of sorter initialization and renewal thereof is passed through, the apparent model of on-line study and more fresh target, thus the precision and the speed that substantially increase target following.

And, the compressed sensing dimension reduction method that implement device provided by the invention have employed based on binary feature expresses the apparent of target, can the deformation of effective expression target, improve anti-blocking and the ability of illumination, thus can robust tracking target, there is the advantage that memory consumption is low and calculated amount is little simultaneously, reach real-time follow-up speed.Therefore there is good using value in actual applications.

Accompanying drawing explanation

Fig. 1 is that the point of stochastic generation of the present invention is to schematic diagram;

Fig. 2 is image block dot matrix trellis diagram exemplary plot of the present invention;

Fig. 3 is feature extracting method schematic diagram of the present invention;

Fig. 4 is Feature Dimension Reduction exemplary plot of the present invention;

Fig. 5 is monotrack method flow diagram of the present invention;

Fig. 6 is the implement device structural representation of monotrack method of the present invention;

Fig. 7 is sample architecture unit process flow diagram of the present invention;

Fig. 8 is model modification unit process flow diagram of the present invention;

Fig. 9 is target tracking unit process flow diagram of the present invention;

Figure 10 and Figure 11 is tracking effect figure of the present invention.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

The invention provides a kind of monotrack method, as shown in Figure 5, suppose the known video V={F be made up of N frame gray scale pedestrian image ₀, F ₁..., F _n-1, the wide height of two field picture is respectively w, h, comprises the steps:

The first step, to get parms

As in Fig. 5 1. shown in, concrete initialization step is as follows:

1) manually, at F ₀target O is obtained in frame ₀the rectangle frame B of initial position ₀=[x ₀, y ₀, w ₀, h ₀], wherein x ₀, y ₀, w ₀, h ₀represent the upper left corner horizontal ordinate of frame respectively, upper left corner ordinate, frame is wide, and frame is high.

2) on the image block I of wide height identical (such as 32x32), L random point is generated to collection wherein represent respectively l point to the horizontal ordinate of first point, the ordinate of first point, the horizontal ordinate of second point, the ordinate of second point.The right generating mode of random point is limited in level or vertical two kinds.As shown in Figure 1, be the right schematic diagram of two random points.Concrete steps are as follows:

A) on image I, grid point set is generated as shown in Figure 2.

B) according to S, can obtain a little to set right to each point two ordinate points additional a random number, i.e. the 4th ordinate respectively $p_4^l=min(h,p_2^l(1+rand))$ With second ordinate $p_2^l=min\{0,p_2^l(1-rand)\},$ Wherein rand represents the random number that [0,1] is interval.Generate point about vertical direction thus to set

C) according to S, can obtain a little to set right to each point two horizontal ordinate points additional a random number, i.e. the 3rd horizontal ordinate respectively $p_3^l=min(w,p_1^l(1+rand))$ With first horizontal ordinate $p_1^l=max(0,p_1^l(1-rand)),$ Wherein rand represents the random number that [0,1] is interval.Generate point about horizontal direction thus to set ${\mathrm{CP}}_h={\left\{\&lsqb;p_1^l,p_2^l,p_3^l,p_2^l\&rsqb;\right\}}_{l=1}^{1024}.$

D) point about vertical and horizontal direction is merged to set notice that the point eliminating repetition is here right, the element number of set CP is L.

3) sparse stochastic matrix A=[α is generated _ij] _{h × L}, ranks are respectively H, L.Wherein the value of line number is 50-300, and optimum value is 100, there will be a known an equiprobability function rand, it generates equiprobably 1,2,3 ..., an element in 2024}.If rand ∈ 1,2,3 ..., 16}, then if rand ∈ 17,18,19 ..., 32}, then otherwise a _ij=0.Sparse stochastic matrix A is used for Feature Dimension Reduction, reduces calculated amount and Noise Resistance Ability.

Second step, sorter initialization and renewal thereof

As in Fig. 5 3., 4. and 5. shown in step, suppose to carry out t (t=0,1 ..., N-1) and secondary iteration renewal, by process t two field picture F _t, iterative process is as follows:

1) training sample set D _tstructure

A) positive sample collection from object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_t^{pos}=\left\{(x,y)\right|x_t-10<x<x_t+10,y_t-10<y<y_t+10\}$ In random extract 50-500, the best is 100 positive samples pictures collection acquisition methods is Pan and Zoom.As shown in Fig. 7 process flow diagram, cycle index N is set as that 100. steps are as follows:

I. the generation formula of positive sample boundary frame:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(1)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii., in the present embodiment, example is operating as follows to carry out 100 times: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (1) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I _i; By I _ibe normalized to the image I that wide height is 32x32 _i.After so carrying out 100 times, generate 100 positive samples pictures set, be designated as

B) negative sample collection ? outer peripheral areas, definitely $Q_t^{neg}=\left\{(x,y)\right|0\&le;x,x\&GreaterEqual;x_t-10,0\&le;y,y\&GreaterEqual;y_t+10\},$ Random acquisition 50-1000 (the best is 100) negative sample collection acquisition methods is translation convergent-divergent.As shown in Fig. 7 process flow diagram, cycle index N is set as 150-500 (the best is 200).Step is as follows:

I. the formula of negative sample bounding box is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(2)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii., in the present embodiment, example is operating as follows to carry out 200 times: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (2) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I; I is normalized to the image I of wide height identical (such as 32x32).After so carrying out 200 times, generate 200 negative sample pictures, be designated as

C) positive negative sample is merged, composing training sample set D _t.Definitely, wherein y _i{-1, the 1} class label representing sample ,-1 represents negative sample to ∈, and 1 represents positive sample.

2) feature extraction, as shown in Figure 3.Extract training sample set D _tin the feature of all sample images, extract sample { I _i, y _ithe step of feature is as follows:

A) initialization sample { I _i, y _ifeature be characteristic length is the element number of CP, i.e. L.

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q).

C) as shown in Figure 4, sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is 50-300 (the best is 100), thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

Structure obtains training sample set D thus _tfeature set

Sorter initialization or renewal, as shown in Figure 8, utilize training sample set D _tfeature set Z _tupgrade sorter model wherein i.e. Renewal model parameter w _t∈ R ^{1 × 101}, R represents real number.Step is as follows:

If d) t=0, initialization w _t∈ R ^{1 × 101}, λ value is 0.0001; Otherwise, perform step b);

E) following iterative step is carried out:

I. from Z _ta middle Stochastic choice k sample, forms subset

Ii. from A _tmiddle searching meets the sample set of certain condition

Iii. calculating parameter value η _t=1/ (λ t).

Iv. first time undated parameter:

$w_{t+\frac{1}{2}}=(1-{\mathrm{\&eta;}}_t\mathrm{\&lambda;})w_t+\frac{\mathrm{\&eta;}t}{k}\underset{\{z,y\}\&Element;A_t^{+}}{\&Sigma;}yz$

V. second time undated parameter:

$w_{t+1}=min\left\{1,\frac{1/\sqrt{\mathrm{\&lambda;}}}{\left|\right|w_{t+\frac{1}{2}}\left|\right|}\right\}{w}_{t+\frac{1}{2}}$

Wherein min represents the minimum value in element, || || represent 2 norms.

F) w is exported _t+1, i.e. model f _t+1.

3rd step, tracking target

As in Fig. 5 6.-9. step, particularly as shown in Fig. 9 process flow diagram.Utilize model f _t+1at F _t+1two field picture carries out target following, and tracking step is as follows:

1) sample set extract.From object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_{t+1}^u=\left\{(x,y)\right|x_t-30<x<x_t+30,y_t-30<y<y_t+30\}$ In random extract 50-500 (the best is 200) positive sample set acquisition methods is Pan and Zoom etc.Step is as follows:

I. the formula of sample boundary frame is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(3)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii. carry out 50-500 following operation, be operating as example as follows to carry out 200 times in the present embodiment: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (3) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _t+1subimage I; I is normalized to the image I of wide height identical (such as 32x32).

Iii. according to step I i, generate 200 samples pictures set, be designated as sample

This set sample frame set is in the picture designated as

2) calculate in the feature of every pictures.Feature extracting method is as shown in Figure 3 identical.Form sample set to be sorted

3) model f is utilized _t+1to U _t+1in all samples classify.Each sample z _i∈ U _t+1produce corresponding degree of confidence:

${\mathrm{Conf}}_i=f_{t+1}\left({\hat{z}}_i\right)=w_{t+1}{\hat{z}}_i$

Wherein degree of confidence is designated as Conf _t+1={ conf ₁, conf ₂..., conf ₂₀₀}

4) according to Conf _t+1, from 10 bounding boxes that middle selection degree of confidence is the highest utilize the object boundary frame B that weighting Meanshift clustering method (DalalN.Findingpeopleinimagesandvideos [D] .InstitutNationalPolytechniquedeGrenoble-INPG, 2006.) generation one is final _t+1=[x _t+1, y _t+1, w _t+1, h _t+1];

5) t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return sorter initialization and the renewal thereof of second step.

As shown in Figure 10, the tracking effect of pedestrian under different video frame, in picture, pedestrian has and changes one's clothes, and turns round, the features such as deformation, and the method that this patent proposes can effectively address these problems.As shown in figure 11, the tracking effect of pedestrian under different video frame, pedestrian has fuzzy, illumination variation, too little, float, the features such as tracking box background is more, and the method that this patent proposes can effectively overcome the above problems.This patent has extremely strong tracking power, has anti-illumination variation, target deformation, apparent change and is subject to the features such as background influence is low.

In sum, a kind of monotrack method provided by the invention is creationary proposes feature extracting method and feature dimension reduction method, the compressed sensing feature extremely strong to Objective extraction robustness, improve the ability to express that target is apparent, then the step of sorter initialization and renewal thereof is passed through, the apparent model of on-line study and more fresh target, thus the precision and the speed that substantially increase target following.

For method for tracking target set forth above, the present invention also provides a kind of implement device of the method, as shown in Figure 6.Image acquiring device, for obtaining a two field picture and carrying out gray processing process and wide high normalized to image from video;

Sorter initialization and updating device thereof, for initialization model and online updating model;

Target tracker, for doing target homing on a new images, makes search result as far as possible consistent with target.

Described image acquiring device comprises random point to generation unit and sparse stochastic matrix generation unit.

Described sorter initialization and updating device thereof comprise samples pictures collection tectonic element, feature extraction unit and model modification unit, and wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing for aligning negative sample image; Model modification unit, the feature samples collection utilizing feature extraction unit to obtain is to upgrade sorter model;

Described target tracker comprises samples pictures collection tectonic element, feature extraction unit and target tracking unit, and wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing, target tracking unit for aligning negative sample image, for classifying to all samples, and therefrom select degree of confidence the most much higher bounding box, and then cluster generates a best object boundary frame.

Further describe the workflow of the implement device unit of a kind of monotrack method of the present invention below,

As shown in Figure 6, first selected i-th two field picture of described image acquiring device, obtains target O ₀the rectangle frame B of initial position ₀=[x ₀, y ₀, w ₀, h ₀], represent the upper left corner horizontal ordinate of frame respectively, upper left corner ordinate, frame is wide, and frame is high;

At wide height be 32x32 image block I on, generate L random point to collection wherein represent respectively l point to the horizontal ordinate of first point, the ordinate of first point, the horizontal ordinate of second point, the ordinate of second point, the right generating mode of random point is limited in level or vertical two kinds, specific as follows:

Image I generates grid point set $S={\left\{\&lsqb;p_1^l,p_2^l\&rsqb;\right\}}_{l=1}^{1024},p_1^l,p_2^l\&Element;\{0,1,2,...31\},$

Vertical point, to generation, according to S, can obtain a little to set right to each point two ordinate points additional a random number, i.e. the 4th ordinate respectively $p_4^l=min(h,p_2^l(1+rand))$ With second ordinate $p_2^l=min\{0,p_2^l(1-rand)\},$ Wherein rand represents the random number that [0,1] is interval.Generate point about vertical direction thus to set

Level point, to generation, according to S, can obtain a little to set right to each point two horizontal ordinate points additional a random number, i.e. the 3rd horizontal ordinate respectively $p_3^l=min(w,p_1^l(1+rand))$ With first horizontal ordinate $p_1^l=max(0,p_1^l(1-rand)),$ Wherein rand represents the random number that [0,1] is interval.Generate point about horizontal direction thus to set

Merge point about vertical and horizontal direction to set the point eliminating repetition is right, and the element number of set CP is L;

Generate sparse stochastic matrix A=[a _ij] _{100 × L}, ranks are respectively 100 and L, there will be a known an equiprobability function rand, it generates equiprobably 1,2,3 ..., an element in 2024}, if rand ∈ 1,2,3 ..., 16}, then if rand ∈ 17,18,19 ..., 32}, then otherwise a _ij=0.

Sorter initialization and updating device thereof, as in Fig. 6 2.-4. shown in, comprise samples pictures set structure unit, feature extraction unit and model modification unit.Wherein, described positive and negative samples pictures set tectonic element: be mainly used in constructing positive sample set of sub-images and negative sample set of sub-images from samples pictures.Described feature extraction unit module: the feature extraction carrying out compressed sensing for aligning negative sample image.Described model modification unit: the feature samples collection obtained for utilizing feature extraction unit is to upgrade sorter model.

Suppose to carry out t (t=0,1 ..., N-1) and secondary iteration renewal, by process t two field picture F _t, iterative process is as follows:

1) training sample set structure

A) positive sample collection from object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_t^{pos}=\left\{(x,y)\right|x_t-10<x<x_t+10,y_t-10<y<y_t+10\}$ In random extract 100 positive samples pictures collection acquisition methods is Pan and Zoom.As shown in Fig. 7 process flow diagram, cycle index N is set as 100.

I. the generation formula of positive sample boundary frame:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(1)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii. carry out 80-150 to operate as follows, be operating as example as follows with 100 times in the present embodiment: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (1) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I _i; By I _ibe normalized to the image I that wide height is 32x32 _i.After so carrying out 100 times, generate 100 positive samples pictures set, be designated as

B) negative sample collection ? outer peripheral areas, definitely, $Q_t^{neg}=\left\{(x,y)\right|0\&le;x,x\&GreaterEqual;x_t-10,0\&le;y,y\&GreaterEqual;y_t+10\},$ Random acquisition 100 negative sample collection acquisition methods is translation convergent-divergent.As shown in Fig. 7 process flow diagram, cycle index N is set as that 200. steps are as follows:

I. the formula of negative sample bounding box is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(2)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii. 150-500 operation is as follows carried out, to carry out 200 operations as follows in the present embodiment: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (2) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I; By I _tbe normalized to the image I that wide height is 32x32.After so carrying out 200 times, generate 200 negative sample pictures, be designated as

C) positive negative sample is merged, composing training sample set D _t.Definitely, wherein y _i{-1, the 1} class label representing sample ,-1 represents negative sample to ∈, and 1 represents positive sample.

2) feature extraction; Extract training sample set D _tin the feature of all sample images, extract sample { I _i, y _ithe step of feature is as follows:

A) initialization sample { I _i, y _ifeature be characteristic length is the element number of CP, i.e. L.

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q).

C) as shown in Figure 4, sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is 50-300, dimension preferably 100 in the present embodiment, thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

The training sample constructed thus is designated as

3) model modification unit, utilizes training set Z _tupgrade sorter model wherein i.e. Renewal model parameter w _t∈ R ^{1 × 101}, R represents real number.Step is as follows:

If t=0, initialization w _t∈ R ^{1 × 101}, λ value is 0.0001; Otherwise, perform step b);

Carry out following iterative step:

From Z _ta middle Stochastic choice k sample, forms subset

From A _tmiddle searching meets the sample set of certain condition

Calculating parameter value η _t=1/ (λ t).

Undated parameter for the first time:

$w_{t+\frac{1}{2}}=(1-{\mathrm{\&eta;}}_t\mathrm{\&lambda;})w_t+\frac{\mathrm{\&eta;}t}{k}\underset{\{z,y\}\&Element;A_t^{+}}{\&Sigma;}yz$

Second time undated parameter:

$w_{t+1}=min\left\{1,\frac{1/\sqrt{\mathrm{\&lambda;}}}{\left|\right|w_{t+\frac{1}{2}}\left|\right|}\right\}{w}_{t+\frac{1}{2}}$

Wherein min represents the minimum value in element, || || represent 2 norms.

Export w _t+1, i.e. model f _t+1.

This sorter initialization and updating device thereof mainly provide the online updating function to model, achieve in object tracking process, and along with target shape, the change of illumination and size, unceasing study target apparent, to improve the robustness of tracking.

Target tracker, for doing target homing on a new images, make search result as far as possible consistent with target, the feature samples set utilizing model classifiers to align the acquisition of negative sample characteristics of image set constructing module is classified, and obtains best target bezel locations.Described target tracker comprises samples pictures collection tectonic element, feature extraction unit and target tracking unit.As in Fig. 6 5.-7. shown in, wherein, samples pictures collection tectonic element, feature extraction unit are identical with flow process with the method for feature extraction unit with the samples pictures collection tectonic element in sorter initialization and updating device thereof.Its workflow is as follows:

1) sample set extract.From object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_{t+1}^u=\left\{(x,y)\right|x_t-30<x<x_t+30,y_t-30<y<y_t+30\}$ In random extract 200 positive samples pictures collection acquisition methods is Pan and Zoom etc.Step is as follows:

I. the formula of sample boundary frame is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(3)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20] .x, y span is

Ii. example is operating as follows to carry out 200 times: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (3) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _t+1subimage I; I is normalized to the image I that wide height is 32x32.

Iii. according to step I i, generate 200 samples pictures set, be designated as sample set sample frame set is in the picture designated as

2) calculate in the feature of every pictures.Feature extracting method is as shown in Figure 3 identical.Form sample set to be sorted

3) model f is utilized _t+1to U _t+1in all samples classify.Each sample z _i∈ U _t+1produce corresponding degree of confidence:

${\mathrm{Conf}}_i=f_{t+1}\left({\hat{z}}_i\right)=w_{t+1}{\hat{z}}_i$

Wherein degree of confidence is designated as Conf _t+1={ conf ₁, conf ₂..., conf ₂₀₀}

4) according to Conf _t+1, from 10 bounding boxes that middle selection degree of confidence is the highest weighting Meanshift method is utilized to generate a final object boundary frame B _t+1=[x _t+1, y _t+1, w _t+1, h _t+1];

5) t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return the model modification unit in sorter initialization and updating device thereof.

To sum up, the compressed sensing dimension reduction method that implement device provided by the invention have employed based on binary feature expresses the apparent of target, can the deformation of effective expression target, improve anti-blocking and the ability of illumination, thus can robust tracking target, there is the advantage that memory consumption is low and calculated amount is little simultaneously, reach real-time follow-up speed.Therefore there is good using value in actual applications.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1. a monotrack method, comprising:

The first step, initiation parameter, that is, at F ₀target O is obtained in frame ₀the rectangle frame B of initial position ₀=[x ₀, y ₀, w ₀, h ₀], represent the upper left corner horizontal ordinate of frame respectively, upper left corner ordinate, frame is wide, and frame is high; On the image block I that wide height is identical, generate L random point to collection wherein represent respectively l point to the horizontal ordinate of first point, the ordinate of first point, the horizontal ordinate of second point, the ordinate of second point, the right generating mode of random point is limited in level or vertical two kinds; Generate sparse stochastic matrix A, for Feature Dimension Reduction;

Second step, sorter initialization and renewal thereof, carry out t (t=0,1 ..., N-1) and secondary iteration renewal, by process t two field picture F _t, comprise training sample set D _tstructure, feature extraction and model modification three processing procedures; Wherein, described feature extraction is for extracting described training sample set D _tin all sample image { I _i, y _ifeature, concrete steps are as follows:

A) initialization sample image { I _i, y _ifeature be characteristic length is L;

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein, I _irepresent sample image; y _i∈-1, the 1} class label representing sample ,-1 represents negative sample, and 1 represents positive sample; I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q);

C) sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is 50-300, thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

Obtain training sample set D _tfeature set

3rd step, target following, utilizes model f _t+1at F _t+1two field picture carries out target following, and tracking step comprises: forecast sample constructs, feature extraction, sample classification, select degree of confidence the most much higher several sample, generate a final object boundary frame, output tracking frame, t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return second step.

2. monotrack method as claimed in claim 1, is characterized in that, the sparse stochastic matrix A=of described generation [α _ij] _{h × K}, line number is H, and value is 50-300, and columns is L, there will be a known an equiprobability function rand, it generates equiprobably 1,2,3 ..., an element in 2024}, if rand ∈ 1,2,3 ..., 16}, then if rand ∈ 17,18,19 ..., 32}, then otherwise, a _ij=0.

3. monotrack method as claimed in claim 1 or 2, is characterized in that, described training sample set D _tstructure comprises the steps:

A) positive sample collection from object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_t^{pos}=\left\{(x,y)\right|x_t-10<x<x_t+10,y_t-10<y<y_t+10\}$ In random extract 50-500 positive samples pictures collection step is as follows:

I. the generation formula of positive sample boundary frame:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(1)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, and span [0,20], x, y span is

Ii. 50-500 operation is as follows carried out: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (1) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I _i; By I _ibe normalized to the image I that wide height is identical _i; After so carrying out 50-500 time, generate 50-500 and open positive samples pictures set;

B) negative sample collection ? outer peripheral areas, definitely $Q_t^{neg}=\left\{(x,y)\right|0\&le;x,x\&GreaterEqual;x_t-10,0\&le;y,y\&GreaterEqual;y_t+10\},$ Random acquisition 50-1000 negative sample collection acquisition methods is translation or convergent-divergent, and step is as follows:

I. the formula of negative sample bounding box is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(2)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, span [0,20]; X, y span is

Ii. 50-1000 operation is as follows carried out: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x', y', w', h'] that formula (2) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _tsubimage I; I being normalized to wide height is identical image I; After so carrying out 50-1000 time, generate 50-1000 and open the set of negative sample picture;

C) positive negative sample is merged, composing training sample set D _t.

4. monotrack method as claimed in claim 1 or 2, it is characterized in that, described model modification is for utilizing training sample set D _tfeature set upgrade sorter model wherein i.e. Renewal model parameter w _t∈ R ^{1 × 101}, R represents real number, and step is as follows:

If a) t=0, initialization w _t∈ R ^{1 × 101}, λ value is 0.0001; Otherwise, perform step b);

B) following iterative step is carried out:

I. from Z _ta middle Stochastic choice k sample, forms subset

Ii. from A _tmiddle searching meets the sample set of certain condition

Iii. calculating parameter value η _t=1/ (λ t);

Iv. first time undated parameter:

$w_{t+\frac{1}{2}}=(1-{\mathrm{\&eta;}}_t\mathrm{\&lambda;})w_t+\frac{\mathrm{\&eta;}t}{k}\underset{\{z,y\}\&Element;A_t^{+}}{\&Sigma;}yz$

V. second time undated parameter:

$w_{t+1}=min\left\{1,\frac{1/\sqrt{\mathrm{\&lambda;}}}{\left|\right|w_{t+\frac{1}{2}}\left|\right|}\right\}{w}_{t+\frac{1}{2}}$

Wherein min represents the minimum value in element, || || represent 2 norms;

C) w is exported _t+1.

5. monotrack method as claimed in claim 1 or 2, is characterized in that, describedly utilizes model f _t+1at F _t+1the step that two field picture carries out target following is as follows:

1) sample set extract, from object boundary frame B _t=[x _t, y _t, w _t, h _t] neighborhood $Q_{t+1}^u=\left\{(x,y)\right|x_t-30<x<x_t+30,y_t-30<y<y_t+30\}$ In random extract 200 positive samples pictures collection acquisition methods is translation or convergent-divergent, and step is as follows:

I. the formula of sample boundary frame is generated:

[x',y',w',h']＝scale[x,y,w _t,h _t]+shift(3)

Wherein scale represents zoom ratio, span [0.8,1.2], and shift represents positive integer side-play amount, span [0,20]; X, y span is

Ii. 200 operations are as follows carried out: in the span of x, y, scale, shift, obtain random value respectively; Then the bounding box [x, y, w, h] that formula (3) calculates sample is substituted into; According to frame [x', y', w', h'] cut-away view as F _t+1subimage I; I is normalized to the image I that wide height is identical;

Iii. according to step I i, 200 samples pictures set are generated;

2) calculate the feature of every pictures, calculate sample { I _i, the method for 1} feature is as follows:

A) initialization sample { I _i, the feature of 1} is characteristic length is L;

B) jth ∈ 1,2 ..., L} component value computing formula as follows:

Wherein I _i(p, q) represents image I _iin the gray-scale pixel values of point (p, q);

C) sparse stochastic matrix A couple is utilized carry out dimensionality reduction, dimension is the line number of matrix A, thus obtains new feature z, and computing formula is as follows:

$z=A\stackrel{\&OverBar;}{z}$

As above, sample set to be sorted is formed

3) model f is utilized _t+1to U _t+1in all samples classify, each sample z _i∈ U _t+1produce corresponding degree of confidence:

${\mathrm{Conf}}_i=f_{t+1}\left({\hat{z}}_i\right)=w_{t+1}{\hat{z}}_i$

Wherein degree of confidence is designated as Conf _t+1={ conf ₁, conf ₂..., conf ₂₀₀;

4) according to Conf _t+1, from the most much higher several bounding box of middle selection degree of confidence this number is preferably 1/20 of sample number, generates a final object boundary frame B _t+1=[x _t+1, y _t+1, w _t+1, h _t+1];

5) t=t+1, if t > is N-1, then terminates to follow the tracks of; Otherwise, return second step.

6. an implement device for monotrack method, comprising:

Image acquiring device, for obtaining a two field picture and carrying out gray processing process and wide high normalized to image from video; Described image acquiring device comprises random point to generation unit and sparse stochastic matrix generation unit;

Sorter initialization and updating device thereof, for initialization model and online updating model;

Target tracker, for doing target homing on a new images, makes search result as far as possible consistent with target.

7. the implement device of monotrack method as claimed in claim 6, it is characterized in that, described sorter initialization and updating device thereof comprise samples pictures collection tectonic element, feature extraction unit and model modification unit, wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing for aligning negative sample image; Model modification unit, the feature samples collection utilizing feature extraction unit to obtain is to upgrade sorter model.

8. the implement device of monotrack method as claimed in claim 6, it is characterized in that, described target tracker comprises samples pictures collection tectonic element, feature extraction unit and target tracking unit, wherein, samples pictures collection tectonic element is used for from samples pictures, construct positive sample set of sub-images and negative sample set of sub-images; Feature extraction unit, carries out the feature extraction of compressed sensing for aligning negative sample image; Target tracking unit, for classifying to all samples, and selects degree of confidence the most much higher bounding box, and then cluster generates a best object boundary frame.